OptNet: Differentiable Optimization as a Layer in Neural Networks

Authors: Brandon Amos, J. Zico Kolter

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present several experimental results that highlight the capabilities of the QP Opt Net layer. Specifically we look at 1) computational efficiency over exiting solvers; 2) the ability to improve upon existing convex problems such as those used in signal denoising; 3) integrating the architecture into an generic deep learning architectures; and 4) performance of our approach on a problem that is challenging for current approaches.
Researcher Affiliation Academia Brandon Amos 1 J. Zico Kolter 1 1School of Computer Science, Carnegie Mellon University. Pittsburgh, PA, USA. Correspondence to: Brandon Amos <bamos@cs.cmu.edu>, J. Zico Kolter <zkolter@cs.cmu.edu>.
Pseudocode No The paper contains mathematical derivations but no explicit pseudocode or algorithm blocks.
Open Source Code Yes The code and data for our experiments are open sourced in the icml2017 branch of https://github.com/locuslab/optnet and our batched QP solver is available as a library at https: //github.com/locuslab/qpth.
Open Datasets No The paper describes using a custom-created dataset for Sudoku ('9000 training puzzles') and generated signals for denoising ('generate piecewise constant signals... and corrupt them with independent Gaussian noise') but does not provide concrete access information (link, DOI, formal citation) for a publicly available or open dataset.
Dataset Splits No The paper mentions 9000 training puzzles and 1000 held-out test puzzles for Sudoku, but does not explicitly specify a validation split or its size.
Hardware Specification Yes All of the Opt Net performance results in this section are run on an unloaded Titan X GPU. Gurobi is run on an unloaded quad-core Intel Core i7-5960X CPU @ 3.00GHz.
Software Dependencies No The paper mentions using the 'Py Torch library' and 'CUBLAS extension' but does not specify their version numbers or any other software dependencies with specific versions.
Experiment Setup Yes We trained these models using ADAM (Kingma & Ba, 2014) to minimize the MSE (which we refer to as loss ) on a dataset we created consisting of 9000 training puzzles, and we then tested the models on 1000 different heldout puzzles.